Amyotrophic Lateral Sclerosis (ALS) is a fatal neurodegenerative disorder characterized by rapid progression and poor prognosis. Early diagnosis remains highly challenging due to nonspecific clinical manifestations, leading to persistently high misdiagnosis rates. This study investigates the feasibility of using resting-state electromyography (EMG) signals for accurate ALS identification. Clinically collected EMG signals vary in length across individuals and muscle regions; further, preprocessing via sliding window methods yields inconsistent numbers of segmented samples, posing challenges for traditional neural networks to adapt to variable-length sequences. To address these issues, we propose a unified time-frequency learning framework called TF-Transformer. Centered on a Transformer-based feature extraction module, the framework treats segmented EMG data as sequence data, enabling efficient processing of variable-length sequences without padding or truncation. It synchronously extracts and fuses time and frequency domain features into a unified EMG representation for ALS identification. Experiments using five-fold cross-validation show that this method effectively utilizes the inherent time-frequency characteristics of resting state electromyographic signals, providing new technical insights and methods for clinical ALS diagnosis.

In recent years, Multimodal Sentiment Analysis (MSA) has attracted growing attention for its ability to interpret human emotions by integrating information across multiple modalities. Multimodal Conversational Aspect-based Sentiment Analysis (MCABSA) extends this research frontier by incorporating multi-party conversational contexts and requiring comprehensive extraction of sentiment elements. MCABSA presents substantial challenges, including the need to understand complex conversational contexts, integrate heterogeneous multimodal signals, and identify causal reasoning at the cognitive level. To address these challenges, we propose a two-stage Full Fine-tuning and LLM Post-processing (FLP) framework. In the first stage, we develop a multimodal caption-enhanced full fine-tuning pipeline that performs structured extraction of sextuples and sentiment flip tuples. The second stage introduces paraphrase-based sextuple verification to identify and filter low-quality sextuples for Panoptic Sentiment Sextuple Extraction (Task-1), while implementing trigger classification with a distribution alignment mechanism to determine trigger types for sentiment flipping and enhance output consistency for Sentiment Flipping Analysis (Task-2). Comprehensive experiments on both MCABSA challenge subtasks demonstrate the effectiveness of our approach, achieving 1st place on Task-1 and 3rd place on Task-2.

Crevasse detection is crucial for glacier and climate research, and provides essential guidance for activities in glacier regions. Ground-penetrating radar(GPR) and machine learning are used to automatically detect crevasse. In this study, a Siam-Gabor-ResNet deep learning framework is proposed to detect crevasse automatically using GPR data. A contrast learning with Siamese network framework is proposed to improve the accuracy of crevasse detection, which aims to increase the feature similarity between crevasses while simultaneously enhancing the feature distinctiveness between crevasse and continuous snow layers. Additionally, a trainable Gabor-ResNet feature extraction module is built by integrating the Gabor filter bank into the ResNet network and used to further reduce the complexity of model training while extracting multi-scale features in a real-time manner. Experiments are performed on the Greenland dataset and the 2015 McMurdo dataset, which illustrate the effectiveness of the proposed method. The average accuracy rate of crevasse detection reaches 94.38%, which can detect the narrowest crevasse (0.6 meters) in two datasets, with an average detection time of only 6.8 milliseconds. Experimental results show that the proposed method can detect crevasse in real-time, automatically, and accurately.

With the explosive popularity of social media, more and more people, including those with depressive symptoms, are starting to express their emotions online through vlogs recently, which makes it important for video-based depression recognition. As video data contains rich acoustical and visual information, the main challenges faced by existing methods include (1) how to accurately mine features associated with depression in massive data and (2) how to effectively fuse various features from different modalities. In this paper, a multi-domain acoustical-visual information fusion network (MDAVIF) is designed to extract depressive spatio-temporal features from image sequences and audios, and an adaptive feature interaction module is proposed to mix these features. Combined with two autoencoders to retain information and prevent overfitting, the proposed method obtains the state-of-the-art result with the precision of 74.25% and the F1-Score of 75.25% when evaluated on the D-vlog dataset.

With the increasing number of depressed patients and the development of computer vision technology, the study of individual automatic depression estimation (ADE) methods based on facial images has attracted much attention in recent years. Most existing works focus on obtaining informative features from the whole images with advancing deep learning models, while the local tiny changes and the fusion of different features have been paid less attention. In this paper, a two-branch predicting model with a elaborate transfomer block NFFT is designed to combine global and local features extracted from whole images and image patches for predicting the depression score precisely. Besides, a classification head is added to guide the regression results for improving accuracy. Experiment results on the AVEC2014 dataset (MAE=5.81, RMSE=7.49) demonstrate that the proposed model outperforms other methods, and the extended experiments on one new dataset (CCPL) are conducted to validate the generalizability and robustness of our model.